Machine Learning-Based Supply Chain Management

Abstract

Historical information features containing supply chain performance metrics (e.g., case fulfillment rate performance metrics) for each of several different time windows are accessed, and then at least a portion of the historical information features are weighted differently for at least a portion of the historical information features according to at least a first criterion, and a training corpus is provided. Subsequently, at least one machine learning model can be trained with the training corpus to generate a machine learning model (or multiple models) configured to predict supply chain performance.

Description

[Technical Field]

[0001] This instruction generally concerns the use of artificial intelligence, and more specifically, the use of machine learning models. [Background technology]

[0002] Supply chain management may involve attempts to understand (and plan accordingly) future needs and the corresponding supplies to meet those needs. Certain key performance indicators may be used to facilitate and / or measure the functioning of the logistics organization.

[0003] Supply chain managers, with or without computer assistance, create metrics to describe anticipated future needs and the corresponding expected and available supply. For example, looking three weeks ahead for a given product, a manager might determine that the demand for that week is X and the available inventory is XY (i.e., somewhat less than the anticipated demand). While this is typically useful information, the applicant determined that such information is not always entirely reliable, and that gaining a sense of how reliable (or unreliable) such conclusions are can itself be a valuable consideration when managing the corresponding supply chain.

[0004] Supply chain management may involve making changes to the supply chain so that its physical elements and configurations function more efficiently and can physically meet future requirements and inventory deliveries. The ability to predict supply chain performance means that the supply chain can be managed more effectively. [Brief explanation of the drawing]

[0005] The above needs, especially when considered in conjunction with the drawings, can be at least partially met by providing machine learning-based supply chain performance forecasts as described in the detailed explanation below. [Figure 1] This figure shows a timeline configured according to various embodiments of this instruction. [Figure 2] This is a schematic diagram configured according to various embodiments of this instruction. [Figure 3] This figure shows an index configured according to various embodiments of this instruction. [Figure 4] This figure shows a timeline configured according to various embodiments of this instruction. [Figure 5] This graph is constructed according to various embodiments of this instruction. [Figure 6] These are figures and graphs showing timelines configured according to various embodiments of this instruction. [Figure 7] This is a schematic diagram configured according to various embodiments of the present invention. [Figure 8] This is a block diagram configured according to various embodiments of this instruction.

[0006] Elements in the figures are shown for brevity and clarity and are not necessarily drawn to scale. For example, the dimensions and / or relative positions of some elements in the figures may be exaggerated relative to others, which is to aid in understanding the various embodiments of this teaching. Also, common elements that are useful or necessary in commercially viable embodiments but are well understood are often not illustrated so as not to obstruct the view of these various embodiments of this teaching. Certain operations and / or steps may be described or illustrated in a particular order of occurrence, but those skilled in the art will understand that such specificity regarding sequence is not actually essential. Terms and expressions used herein have the ordinary technical meanings given to them by those skilled in the art unless otherwise explicitly stated. The word “or” used herein shall be interpreted as a disjunctive rather than conjunctive construction unless otherwise explicitly stated. [Modes for carrying out the invention]

[0007] Generally speaking, these various embodiments allow access to historical information features, each containing a supply chain performance metric (e.g., a case fulfillment rate performance metric) for each of several different time windows, providing a training corpus by differently weighting at least a portion of the historical information features according to at least a first criterion, and then training at least one machine learning model using that training corpus to generate a machine learning model (or more models) configured to predict supply chain performance.

[0008] According to one approach, the first criterion described above includes the temporal proximity of each of the different time windows to its respective target time window (e.g., a future time window, but not limited to one). According to one approach, differently weighting at least some of the historical information features according to at least the first criterion described above includes, in at least some, weighting at least one of the different time windows that is closer to a future time window more highly than another of the different time windows that is further from a future time window.

[0009] According to one approach, the first criterion described above includes inventory week parameters corresponding to each of the different time windows. In this case, differently weighting at least some of the historical information features according to at least the first criterion may include, in at least some respects, giving higher weight to at least one of the different time windows that has inventory week parameters sufficiently similar to those of the future time window than to another of the different time windows that has inventory week parameters that are less similar to those of the future time window.

[0010] Therefore, according to one approach, this teaching allows for a control circuit configured as at least one supply chain performance predictive machine learning model, which is at least partially trained on a training corpus formed by accessing historical information features, each containing a supply chain performance metric for each of several different time windows, and differently weighting at least some of the historical information features according to at least a first criterion.

[0011] If necessary, such a supply chain performance prediction machine learning model may be further configured to calculate at least one supply chain performance metric threshold (for example, by analyzing the historical relationship between the case fulfillment rate metric and the weeks-in-supply metric, at least in part, and subsequently calculating at least one supply chain performance metric threshold as a weeks-in-supply metric threshold that identifies a favorable future case fulfillment rate metric).

[0012] One approach is a computer implementation method for managing a supply chain, Accessing historical information features that include supply chain performance metrics for each of multiple different time windows, To provide a training corpus by differently weighting at least a portion of the historical information features according to at least the first criterion, This involves training a machine learning model using a training corpus to generate a machine learning model configured to predict supply chain performance, and Predicting supply chain performance using machine learning models, Modify supply chain planning data based on predicted supply chain performance, Processing inventory within the supply chain (e.g., moving inventory from one location to another, assembling components or parts, and / or manufacturing a product) A computer-implemented method is provided. Thus, by training a machine learning model, the performance of the supply chain can be predicted more accurately. If this predicted performance is less than the required amount (e.g., less than a threshold or a predetermined threshold), the parameters in the planning data used to implement the supply chain can be changed so that the supply chain can function more effectively. This may lead to, for example, inventory being delivered more reliably and quickly, and with a reduced number of delivery slots. This can save fuel and other resources.

[0013] According to one approach, further, the present teachings allow accessing historical information features each including a supply chain performance metric for each of a plurality of different time windows, at least in part, and inputting information into a supply chain performance prediction machine learning model trained by a training corpus formed by differently weighting at least a part of the historical information features according to at least a first criterion, and subsequently outputting a supply chain performance prediction for a future time window from the supply chain performance prediction machine learning model.

[0014] According to one approach, such a supply chain performance prediction machine learning model can be configured to calculate at least one supply chain performance metric threshold.

[0015] According to another approach, instead of or in combination with the above, such a supply chain performance prediction can include, at least in part, a supply weeks metric that yields a preferred case fulfillment rate metric for a future time window.

[0016] These and other advantages may become more apparent by carefully reading and considering the following detailed description. Referring to the drawings, particularly FIG. 1, an example of an exemplary temporal context is first presented.

[0017] As described herein, this instruction relates to predicting the status / performance of a supply chain in different time windows. For the sake of convenience and without any intention to suggest any limitations on the duration or periodicity of these windows (or even whether all windows are of similar duration), the following explanation assumes that each time window in question is a period of one week (i.e., seven consecutive days). Starting from the "present" 101 (i.e., the current time window), future time windows 102 (in this example, "weeks") can be expressed sequentially and serially as week 0, week 1, etc., up to week N (where "N" is an integer), or thereafter. For many applications, providing 52 weeks is useful and beneficial, but this number can be easily changed as needed. Similarly, working backward from the "present" 101, past time windows 103 (likewise, "weeks" in this example) can be expressed sequentially as needed as lag-0, lag-1, etc., up to lag-N, or earlier ("lag" will be discussed later as appropriate).

[0018] Generally, this instruction concerns making supply chain forecasts for each of several future time windows (in this case, weeks) and evaluating how confident one feels in those forecasts. The confidence metric can be used in various ways.

[0019] Referring to Figure 2, we will describe an approach to estimating / forecasting key performance indicators in a supply chain. In this example, for illustrative purposes, the key performance indicator is the case fulfillment rate (CFR). The case fulfillment rate represents the difference between the number of products ordered by a customer and the number of products actually shipped. For example, if a customer orders 100 cases of an item but only 90 cases are actually shipped, the case fulfillment rate for that item would be 90%. If future CFR risks (i.e., the expected undesirable discrepancies between orders and fulfillment) can be accurately predicted, a forward-thinking organization can more effectively prevent the materialization of those risks.

[0020] The approach shown in Figure 2 involves accessing data 201, which may include historical plan data for all so-called snapshot weeks and all so-called accounting weeks over the relevant historical period (e.g., the past year). A snapshot week is the week in which the planner made the corresponding plan regarding the expected fulfillment of orders, and an accounting week is a future week to which the above plan applies (i.e., a future week compared to the corresponding snapshot week).

[0021] The difference between an accounting week and its corresponding week (for example, a snapshot week, though not limited to one) is called a lag.

[0022] In this example, we assume that the planner creates / adjusts plans weekly. These plans are formed as functions of demand, production, and inventory. As will be detailed later, this approach compares the planner's past forecasts with actual achievements (i.e., results) to facilitate adjustments to the current plan.

[0023] One approach, as illustrated, may involve preprocessing 202 for some or all of the data 201. This preprocessing 202 may vary depending on the needs and / or opportunities presented by a given use case. Generally, the data 201 may be cleaned, for example, to remove irrelevant or duplicate content, correct syntax and / or formatting, etc.

[0024] One approach is to include the generation of leading indicators. Leading indicators capture the inertia of the corresponding supply chain. For example, if new inventory arrivals cannot be shipped in the same week and cannot be shipped until the following week or later, the previous week's inventory may be a better indicator than the current week's inventory. Preprocessing 202 can create features from, for example, the past four weeks. These features allow downstream machine learning models to learn the upward or downward trend leading up to the current week. As an example, we consider Weeks of Supply (WOS) (details below), demand, production, and CFR as the four variables. Useful corresponding machine learning features may be as follows: WOS wk-3 、WOS wk-3 、WOS wk-2 、WOS wk-1 、WOS wk 、Demand wk-4 、Demand wk-3 、Demand wk-2 、Demand wk-1 、Demand wk 、Production wk-4 、Production wk-3 、Production wk-2 、Production wk-1 、Production wk 、CFR wk-4 、CFR wk-3 、CFR wk-2 、CFR wk-1

[0025] Subsequently, a separate data subset 203 for a single given product (e.g., corresponding to a minimum stock keeping unit (SKU) number) for each of the associated lag windows is extracted and formed from the preceding (optionally preprocessed) data 201. Specifically, data subsets 203 can be formed for each of lag-0, lag-1, lag-2, etc., continuously up to lag-N.

[0026] Subsequently, using the data subset 203 described above, a training data set (i.e., a training corpus for the corresponding machine learning model) 204 is formed. Generally, in this example, each training data set 204 is formed from two or three of the data subsets 203. For example, a training data set labeled "training data lag-1" is formed by combining the data subsets of lag-0, lag-1, and lag-2, and a training data set labeled "training data lag-3" is formed by combining the data subsets of lag-2, lag-3, and lag-4. FIG. 2 shows that, in this example, the first and last training data sets are formed using only two of the data subsets 203 continuously.

[0027] As shown in reference numeral 205, this approach can subsequently provide data prioritization. Generally, this activity serves to assign higher priority to certain data points (for example, assigning higher priority to more recent data points than to older data points).

[0028] Furthermore, as indicated by reference numeral 206, this approach can provide feature prioritization. In this example, this prioritization utilizes the Weeks of Supply (WOS) metric, which measures the relationship between inventory and demand, specifically how many weeks of demand a given inventory covers.

[0029] For example, if there is an inventory of 100 cases at the end of the current weekend, and demand for the following week is 40 cases, and demand for the week after that is 60 cases, then the WOS (Will-Own-Storage) would be 2. However, if the demand for the same 100 cases each week for the following consecutive weeks is 20 cases, then the same 100 cases would correspond to a WOS value of 5 (note that the applicant believes that by utilizing the relationship between past WOS and past CFR when forecasting future CFR, a more sophisticated and accurate model can be created than by simply using the relationship between past CFR and inventory).

[0030] For example, the WOS threshold required to achieve a 95% case fulfillment rate can be calculated based on historical data. One approach suggests that a WOS close to this threshold is close to a 95% case fulfillment rate, and vice versa. The case fulfillment rate for future weeks can be calculated based on this proximity to the threshold. The accuracy of the case fulfillment rate calculated using this method is higher when the WOS is close to the threshold, and lower as the WOS moves further away from such a threshold.

[0031] Based on the above, data from more recent weeks can be given a higher weight than data from more distant weeks. Thresholds such as 95%, 80%, and 65% can be calculated. Three case satisfaction rates can be calculated based on proximity to these thresholds. These case satisfaction rates can be assigned weights based on their relative proximity to the corresponding thresholds. For example, if the predicted case satisfaction rate is 72% based on the 95% threshold, that rate is considered unreliable, and the weight given to that rate may be zero or close to zero. On the other hand, if the case satisfaction rate is 81% based on the 80% threshold, the result is considered more reliable, and a higher weight is assigned to that rate.

[0032] The training dataset 204 for each (arbitrarily prioritized) lag window of reference codes is then used to train a machine learning model. To clarify, as mentioned above, this is done per product, so there may be a separate trained machine learning model for each product in the entire inventory, which may consist of hundreds of thousands of individual products.

[0033] Those skilled in the art understand that machine learning is a branch of artificial intelligence. Machine learning typically uses learning algorithms such as Bayesian networks, decision trees, and nearest neighbors, and can operate in a supervised or unsupervised manner as needed. Deep learning (also known as hierarchical learning, deep neural learning, or deep structured learning) is a subset of machine learning, which uses learnable networks (often supervised, where the data consists of pairs (such as input data and labels) and aims to learn a mapping between the input data and the corresponding labels) from unstructured and / or unlabeled data, at least initially. Architectures of deep learning include deep neural networks, deep belief networks, recurrent neural networks, and convolutional neural networks. Many machine learning algorithms build so-called "models" based on sample data known as training data or training corpora, and make predictions or decisions without being explicitly programmed as such.

[0034] In this example, we assume that the machine learning model includes an N-regression model. An N-regression model is a type of regression model that involves predicting a numerical target variable based on multiple input features. "N" refers to the number of input features used to make predictions. In N-regression, the model learns the mathematical relationship between the input features and the target variable by minimizing the difference between the predicted and actual values ​​of the target variable. The goal is to develop a model that can accurately predict the target variable based on the input features of a new instance.

[0035] As indicated by reference numeral 208, each trained machine learning model then outputs a predicted case satisfaction rate for the corresponding future week (e.g., week 0, week 1, week 2, etc.). Such an approach can be performed frequently or infrequently (and regularly, for example, weekly, or randomly), depending on the given application.

[0036] Therefore, it is understood and recognized that this instruction provides an effective method for making effective use of past prediction results and for better evaluating / establishing future predictions.

[0037] A more detailed example of this point is given below. Please note that the details of this example are for illustrative purposes only and are not intended to imply any restrictions on the implementation of this instruction.

[0038] Referring to Figure 3, the three geometric icons help indicate past time windows in which the corresponding predictions of case satisfaction rates were accurate at a given percentage level. Thus, in the following example, rectangle 301 indicates that the resulting actual case satisfaction rate was accurate at 97% or higher. Circle 302 indicates that the resulting actual case satisfaction rate was accurate in the range of 90% to less than 97%. Triangle 303 indicates that the resulting actual case satisfaction rate was accurate at less than 90%.

[0039] Figure 4 shows a series of timeframes, each beginning with a specific past snapshot week 401. For example, the first row 402 begins with snapshot week 401 of week 4 of 2021, and the third row 403 begins with snapshot week 401 of week 12 of 2021. Each row is followed by the corresponding snapshot week 401, with lag weeks moving to the right. One of the geometric icons described above helps to indicate the accuracy of the predicted case fulfillment rate for each particular week in this example. For example, in the first row 402, the week corresponding to lag=6 (i.e., week 10 of 2021) is shown with a square geometric icon indicating that the accuracy result is above the 97th percentile.

[0040] In this example, the current focus is on weeks corresponding to lag=4 (represented collectively by the bounding box indicated by reference numeral 405). Figure 5 shows a graph of past case satisfaction rates corresponding to weeks with lag=4, on Graph 500, defined by the first variable on the X-axis and the second variable on the Y-axis. For clarity and brevity, only two variables / two dimensions are shown here. In typical application settings, the machine learning model used may consider many more variables.

[0041] Figure 6 shows timeline 601, which begins with snapshot week 401 (in this example, week 45 of 2021; the timeline shown in Figure 4 ends with week 44 of the same year, immediately preceding the current week). Therefore, in this example, the historical data under consideration begins with week 4 of 2021 and ends with week 44 of the same year, which is immediately preceding the current week, serving as a snapshot week for predicting future case fulfillment rates on a weekly basis with a forward-looking perspective.

[0042] As mentioned above, this example focuses particularly on the history related to lag=4. This is because the goal in this example is to establish a case fulfillment rate forecast for a future week four weeks ahead from the snapshot week (week 4 is indicated by reference number 602). As shown in Graph 500 (shown in Figure 5), the goal is to determine how accurately the predicted case fulfillment rate is positioned.

[0043] Figure 7 illustrates the estimation of the confidence level of the predicted case fulfillment rate using unsupervised learning as described herein. In this example, the features of the data points used may include demand, production, inventory, transportation, etc. In this example, past weeks (see Figure 4) can be divided into n clusters based on feature similarity. Based on which cluster future weeks resemble, the probability of service delivery failures in future weeks can be calculated (note that lag is not used as a feature in this example).

[0044] In Figure 7, each geometric shape represents a past week, along with an indicator of the accuracy of the corresponding case fulfillment rate prediction for that week. A past week may consist of multiple input features and one output (i.e., case fulfillment rate). The week is divided into n clusters (typically indicated by reference numeral 701) using an unsupervised learning method as described herein. The cluster members for future weeks can be calculated, and then the failure probability can be calculated based on how many service delivery failures (failures are represented by triangular geometric icons) occurred in any given cluster.

[0045] For example, if a future week (indicated by a star icon and reference numeral 702) belongs to cluster 2 (indicated by reference numeral 703), we can confidently predict that the probability of a service outage is 100%. On the other hand, if a future week (indicated by an X-shaped icon and reference numeral 704) belongs to cluster 1 (indicated by reference numeral 705), the probability of a service outage is only 22%.

[0046] Various examples of the enabling device 800 to support the above teaching will be explained with reference to Figure 8.

[0047] For illustrative purposes, we hereby assume that one or more optimal control circuits 801 perform the operations, steps, and / or functions described herein. Being a “circuit,” a control circuit 801 has a structure comprising at least one (typically many) conductive paths (e.g., paths composed of a conductive metal such as copper or silver) such that these paths transmit electricity in an orderly manner and also comprises typically corresponding electronic components (both passive components (such as resistors and capacitors) and active components (such as any of the various semiconductor-based devices) as appropriate) such that the circuit can realize the control surface of this teaching.

[0048] Such a control circuit 801 may comprise a fixed-purpose hardwired hardware platform (including, but not limited to, application-specific integrated circuits (ASICs) (integrated circuits designed for a specific purpose and not intended for general use), field-programmable gate arrays (FPGAs), etc.) or a partially or fully programmable hardware platform (including, but not limited to, microcontrollers, microprocessors, etc.). The architectural options for these structures are well known and understood in the art and do not require further detailed explanation here. The control circuit 801 is configured to perform one or more of the steps, operations, and / or functions described herein (for example, by using corresponding programming that will be well understood by those skilled in the art).

[0049] In this example, the control circuit 801 is operably connected to the memory 802. The memory 802 may be integrated with the control circuit 801 as desired, or it may be physically separated from the control circuit 801 (in whole or in part). The memory 802 may be local to the control circuit 801 (e.g., if they share a common circuit board, chassis, power supply, and / or housing), or it may be partially or completely remote to the control circuit 801 (e.g., if the memory 802 is typically physically located in a different facility, metropolitan area, or country from the control circuit 801). It is also understood that the memory 802 may comprise multiple physically separated memories that store the relevant information corresponding to this teaching as a whole.

[0050] In addition to the aforementioned historical data, the memory 802 can, for example, nonvolatilize computer instructions and machine learning models that, when executed by the control circuit 801, cause the control circuit 801 to operate as described herein (as used herein, “nonvolatilize” refers to the non-temporary state of the stored contents (and therefore, except when the stored contents are merely signals or waves), and includes both nonvolatility memory (such as read-only memory (ROM)) and volatile memory (such as dynamic random-access memory (DRAM))).

[0051] In this example, the control circuit 801 is also operablely connected to the user interface 803. This user interface 803 is equipped with one of various user input mechanisms (keyboard and keypad, cursor control device, touch sensor display, voice recognition interface, gesture recognition interface, etc., but not limited to these) and / or user output mechanisms (display device, sound transducer, printer, etc., but not limited to these), making it easy to receive information and / or instructions from the user, and / or provide information to the user. With this configuration, the information output by the above process can be presented to the user, thereby allowing, for example, the evaluation of the reliability of a given case satisfaction rate prediction for a given future period.

[0052] As one optional approach, the control circuit 801 may also be operablely connected to the network interface 804. This configuration allows the control circuit 801 to communicate with other elements (whether inside or outside the device 800) via the network interface 804. In particular, the control circuit 801 can communicate with one or more remote resources 806 (e.g., a server providing some or all of the aforementioned historical data) and / or one or more remote user interfaces 807 (which, for example, can transmit the aforementioned prediction results to one or more remote users) via one or more intervening networks 805 (such as the known Internet). Network interfaces, including both wireless and wired platforms, are well understood in this art and require no further explanation.

[0053] Those skilled in the art will recognize that a wide range of modifications, changes, and combinations can be made with respect to the embodiments described above without departing from the scope of these teachings. For example, these teachings can be beneficially applied to other key performance metrics. Thus, such modifications, changes, and combinations are also considered to be within the scope of the concept of the present invention.

[0054] The following numbered clauses provide further examples.

[0055] 1. A computer implementation method for managing a supply chain, Accessing historical information features that include supply chain performance metrics for each of multiple different time windows, To provide a training corpus by differently weighting at least a portion of the historical information features according to at least the first criterion, This involves training a machine learning model using a training corpus to generate a machine learning model configured to predict supply chain performance, and Predicting supply chain performance using machine learning models, Modify supply chain planning data based on predicted supply chain performance, Processing inventory within the supply chain and A method that includes this.

[0056] 2. The computer implementation method described in Clause 1, wherein the supply chain performance indicators include case fulfillment performance indicators.

[0057] 3. A computer implementation method as described in Clause 1 or Clause 2, wherein the first criterion includes the temporal proximity of different time windows to each target time window.

[0058] 4. Any computer implementation method described in the preceding clause, wherein the target time window includes a future time window.

[0059] 5. A computer implementation method according to Clause 4, wherein weighting at least a portion of the historical information features differently in accordance with at least the first criterion includes, in at least a portion, weighting at least one of the different time windows that is closer to the future time window more highly than another of the different time windows that is further from the future time window.

[0060] 6. Any computer implementation method described in the preceding clause, wherein the first criterion includes an inventory week parameter corresponding to each of the different time windows.

[0061] 7. A computer implementation method according to Clause 6, wherein differently weighting at least a portion of the historical information features in accordance with at least the first criterion includes, in at least a portion, weighting more highly one of different time windows that has an inventory week parameter sufficiently similar to the inventory week parameter of a future time window than another of different time windows that has an inventory week parameter that is less similar to the inventory week parameter of a future time window.

[0062] 8. A device, Memory and A control circuit operably connected to memory and configured as a supply chain performance prediction machine learning model, wherein the supply chain performance prediction machine learning model is trained with a training corpus formed by accessing historical information features, each containing a supply chain performance metric for each of several different time windows, and weighting at least a portion of the historical information features differently according to at least a first criterion, and the control circuit We use machine learning models to predict supply chain performance. Modify supply chain planning data based on predicted supply chain performance. Processing inventory within the supply chain The control circuit is further configured as follows: A device equipped with the following features.

[0063] 9. The apparatus described in Clause 8, whose supply chain performance indicators include case fulfillment performance indicators.

[0064] 10. The apparatus described in Clause 8 or Clause 9, wherein the first criterion includes the temporal proximity of different time windows to each target time window.

[0065] 11. The apparatus described in Clause 10, wherein the target time window includes a future time window.

[0066] 12. The apparatus according to Clause 11, wherein differently weighting at least a portion of the historical information features in accordance with at least the first criterion includes, in at least a portion, weighting at least one of the different time windows that is closer to the future time window more highly than another of the different time windows that is further from the future time window.

[0067] 13. The apparatus described in any of Clauses 8 to 12, wherein the first criterion includes an inventory week parameter corresponding to each of the different time windows.

[0068] 14. The apparatus according to Clause 13, wherein differently weighting at least a portion of the historical information features in accordance with at least the first criterion includes, in at least a portion, weighting more highly one of different time windows that has an inventory week parameter sufficiently similar to the inventory week parameter of a future time window than another of different time windows that has an inventory week parameter that is less similar to the inventory week parameter of a future time window.

[0069] 15. The apparatus described in any of Clauses 8 to 14, wherein the supply chain performance prediction machine learning model is configured to calculate at least one supply chain performance metric threshold.

[0070] 16. The apparatus according to Clause 15, wherein the supply chain performance prediction machine learning model is configured, at least in part, to calculate at least one supply chain performance metric threshold by analyzing the historical relationship between the case fulfillment rate metric and the weeks of supply metric.

[0071] 17. The apparatus according to Clause 16, wherein the supply chain performance prediction machine learning model is configured to calculate at least one supply chain performance indicator threshold as a supply weeks metric threshold that identifies a favorable future case fulfillment rate metric.

[0072] 18. A method, The method involves inputting information into a supply chain performance prediction machine learning model, wherein the supply chain performance prediction machine learning model is trained using a training corpus formed by accessing historical information features, each containing a supply chain performance metric for each of several different time windows, and weighting at least a portion of the historical information features differently according to at least a first criterion. To output supply chain performance predictions for future time windows from a supply chain performance prediction machine learning model, Modifying supply chain planning data based on supply chain performance forecasts, Processing inventory within the supply chain and A method that includes this.

[0073] 19. The method according to clause 18, wherein the supply chain performance prediction machine learning model is configured to calculate at least one supply chain performance metric threshold.

[0074] 20. The method according to Clause 19, wherein the supply chain performance forecast includes a supply weeks metric that yields a favorable case fulfillment rate metric for future time windows, at least in part.

Claims

1. A computer implementation method for training a machine learning model to predict supply chain performance, Accessing historical information features that include supply chain performance metrics for each of multiple different time windows, To provide a training corpus by differently weighting at least a portion of the historical information features according to at least one criterion, The process involves training a machine learning model using the aforementioned training corpus to generate a machine learning model configured to predict supply chain performance. Computer implementation methods, including those mentioned above.

2. The computer implementation method according to claim 1, wherein the supply chain performance indicator includes a case fulfillment rate performance indicator.

3. The computer implementation method according to claim 1, wherein the first criterion includes the temporal proximity of each of the different time windows to the respective target time window.

4. The computer implementation method according to claim 3, wherein the target time window includes a future time window.

5. The computer implementation method according to claim 4, wherein differently weighting at least a portion of the historical information features in accordance with at least a first criterion includes, in at least a portion, weighting at least one of the different time windows that is closer to the future time window more highly than another of the different time windows that is further from the future time window.

6. The computer implementation method according to claim 1, wherein the first criterion includes an inventory week parameter corresponding to each of the different time windows.

7. The computer implementation method according to claim 6, wherein differently weighting at least a portion of the historical information features in accordance with at least a first criterion includes, in at least a portion, giving a higher weight to at least one of the different time windows having an inventory week parameter that is sufficiently similar to the inventory week parameter of a future time window than to another of the different time windows having an inventory week parameter that is less similar to the inventory week parameter of a future time window.

8. It is a device, Memory and A control circuit operably connected to the memory and configured as a supply chain performance prediction machine learning model, wherein the supply chain performance prediction machine learning model is trained with a training corpus formed by accessing historical information features, each containing a supply chain performance indicator for each of a plurality of different time windows, and weighting at least a portion of the historical information features differently according to at least a first criterion. A device equipped with the following features.

9. The apparatus according to claim 8, wherein the supply chain performance indicator includes a case fulfillment rate performance indicator.

10. The apparatus according to claim 8, wherein the first criterion includes the temporal proximity of each of the different time windows to a target time window.

11. The apparatus according to claim 10, wherein the target time window includes a future time window.

12. The apparatus according to claim 11, wherein differently weighting at least a portion of the historical information features in accordance with at least a first criterion includes, in at least a portion, weighting at least one of the different time windows that is closer to the future time window more highly than another of the different time windows that is further from the future time window.

13. The apparatus according to claim 8, wherein the first criterion includes an inventory week parameter corresponding to each of the different time windows.

14. Apparatus according to claim 13, wherein differently weighting at least a portion of the historical information features in accordance with at least a first criterion includes, in at least a portion, weighting more highly one of the different time windows that has an inventory week parameter sufficiently similar to the inventory week parameter of a future time window than another of the different time windows that has an inventory week parameter less similar to the inventory week parameter of a future time window.

15. The apparatus according to claim 8, wherein the supply chain performance prediction machine learning model is configured to calculate at least one supply chain performance indicator threshold.

16. The apparatus according to claim 15, wherein the supply chain performance prediction machine learning model is configured to calculate at least one supply chain performance indicator threshold by analyzing the historical relationship between the case fulfillment rate metric and the supply week metric, at least in part.

17. The apparatus according to claim 16, wherein the supply chain performance prediction machine learning model is configured to calculate the at least one supply chain performance indicator threshold as a supply week metric threshold that identifies a preferred future case fulfillment rate metric.

18. It is a method, The method involves inputting information into a supply chain performance prediction machine learning model, wherein the supply chain performance prediction machine learning model is trained with a training corpus formed by accessing historical information features, each containing a supply chain performance indicator for each of several different time windows, and weighting at least a portion of the historical information features differently according to at least a first criterion. The machine learning model for predicting supply chain performance outputs supply chain performance predictions for future time windows. Methods that include...

19. The method according to claim 18, wherein the supply chain performance prediction machine learning model is configured to calculate at least one supply chain performance metric threshold.

20. The method according to claim 19, wherein the supply chain performance forecast includes, in at least part, a supply week metric that yields a favorable case fulfillment metric for a future time window.

Images